Support ControlNet for Qwen-Image (#12215)

* support qwen-image-cn-union


---------

Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: YiYi Xu <yixu310@gmail.com>

docs/source/en/api/pipelines/qwenimage.md

@@ -120,6 +120,10 @@ The `guidance_scale` parameter in the pipeline is there to support future guidan
   - all
   - __call__
 
+## QwenImaggeControlNetPipeline
+  - all
+  - __call__
+
 ## QwenImagePipelineOutput
 
 [[autodoc]] pipelines.qwenimage.pipeline_output.QwenImagePipelineOutput

src/diffusers/__init__.py

@@ -218,6 +218,8 @@ else:
             "OmniGenTransformer2DModel",
             "PixArtTransformer2DModel",
             "PriorTransformer",
+            "QwenImageControlNetModel",
+            "QwenImageMultiControlNetModel",
             "QwenImageTransformer2DModel",
             "SanaControlNetModel",
             "SanaTransformer2DModel",
@@ -491,6 +493,7 @@ else:
             "PixArtAlphaPipeline",
             "PixArtSigmaPAGPipeline",
             "PixArtSigmaPipeline",
+            "QwenImageControlNetPipeline",
             "QwenImageEditPipeline",
             "QwenImageImg2ImgPipeline",
             "QwenImageInpaintPipeline",
@@ -885,6 +888,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             OmniGenTransformer2DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
+            QwenImageControlNetModel,
+            QwenImageMultiControlNetModel,
             QwenImageTransformer2DModel,
             SanaControlNetModel,
             SanaTransformer2DModel,
@@ -1128,6 +1133,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PixArtAlphaPipeline,
             PixArtSigmaPAGPipeline,
             PixArtSigmaPipeline,
+            QwenImageControlNetPipeline,
             QwenImageEditPipeline,
             QwenImageImg2ImgPipeline,
             QwenImageInpaintPipeline,

src/diffusers/models/__init__.py

@@ -52,6 +52,10 @@ if is_torch_available():
         "HunyuanDiT2DControlNetModel",
         "HunyuanDiT2DMultiControlNetModel",
     ]
+    _import_structure["controlnets.controlnet_qwenimage"] = [
+        "QwenImageControlNetModel",
+        "QwenImageMultiControlNetModel",
+    ]
     _import_structure["controlnets.controlnet_sana"] = ["SanaControlNetModel"]
     _import_structure["controlnets.controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
     _import_structure["controlnets.controlnet_sparsectrl"] = ["SparseControlNetModel"]
@@ -148,6 +152,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiT2DMultiControlNetModel,
             MultiControlNetModel,
             MultiControlNetUnionModel,
+            QwenImageControlNetModel,
+            QwenImageMultiControlNetModel,
             SanaControlNetModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,

src/diffusers/models/controlnets/__init__.py

@@ -9,6 +9,7 @@ if is_torch_available():
         HunyuanDiT2DControlNetModel,
         HunyuanDiT2DMultiControlNetModel,
     )
+    from .controlnet_qwenimage import QwenImageControlNetModel, QwenImageMultiControlNetModel
     from .controlnet_sana import SanaControlNetModel
     from .controlnet_sd3 import SD3ControlNetModel, SD3ControlNetOutput, SD3MultiControlNetModel
     from .controlnet_sparsectrl import (

src/diffusers/models/controlnets/controlnet_qwenimage.py

@@ -0,0 +1,359 @@
+# Copyright 2025 Black Forest Labs, The HuggingFace Team and The InstantX Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
+from ..attention_processor import AttentionProcessor
+from ..cache_utils import CacheMixin
+from ..controlnets.controlnet import zero_module
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..transformers.transformer_qwenimage import (
+    QwenEmbedRope,
+    QwenImageTransformerBlock,
+    QwenTimestepProjEmbeddings,
+    RMSNorm,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class QwenImageControlNetOutput(BaseOutput):
+    controlnet_block_samples: Tuple[torch.Tensor]
+
+
+class QwenImageControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 64,
+        out_channels: Optional[int] = 16,
+        num_layers: int = 60,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        joint_attention_dim: int = 3584,
+        axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
+        extra_condition_channels: int = 0,  # for controlnet-inpainting
+    ):
+        super().__init__()
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pos_embed = QwenEmbedRope(theta=10000, axes_dim=list(axes_dims_rope), scale_rope=True)
+
+        self.time_text_embed = QwenTimestepProjEmbeddings(embedding_dim=self.inner_dim)
+
+        self.txt_norm = RMSNorm(joint_attention_dim, eps=1e-6)
+
+        self.img_in = nn.Linear(in_channels, self.inner_dim)
+        self.txt_in = nn.Linear(joint_attention_dim, self.inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                QwenImageTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # controlnet_blocks
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            self.controlnet_blocks.append(zero_module(nn.Linear(self.inner_dim, self.inner_dim)))
+        self.controlnet_x_embedder = zero_module(
+            torch.nn.Linear(in_channels + extra_condition_channels, self.inner_dim)
+        )
+
+        self.gradient_checkpointing = False
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self):
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    @classmethod
+    def from_transformer(
+        cls,
+        transformer,
+        num_layers: int = 5,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        load_weights_from_transformer=True,
+        extra_condition_channels: int = 0,
+    ):
+        config = dict(transformer.config)
+        config["num_layers"] = num_layers
+        config["attention_head_dim"] = attention_head_dim
+        config["num_attention_heads"] = num_attention_heads
+        config["extra_condition_channels"] = extra_condition_channels
+
+        controlnet = cls.from_config(config)
+
+        if load_weights_from_transformer:
+            controlnet.pos_embed.load_state_dict(transformer.pos_embed.state_dict())
+            controlnet.time_text_embed.load_state_dict(transformer.time_text_embed.state_dict())
+            controlnet.img_in.load_state_dict(transformer.img_in.state_dict())
+            controlnet.txt_in.load_state_dict(transformer.txt_in.state_dict())
+            controlnet.transformer_blocks.load_state_dict(transformer.transformer_blocks.state_dict(), strict=False)
+            controlnet.controlnet_x_embedder = zero_module(controlnet.controlnet_x_embedder)
+
+        return controlnet
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        controlnet_cond: torch.Tensor,
+        conditioning_scale: float = 1.0,
+        encoder_hidden_states: torch.Tensor = None,
+        encoder_hidden_states_mask: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        img_shapes: Optional[List[Tuple[int, int, int]]] = None,
+        txt_seq_lens: Optional[List[int]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
+        """
+        The [`FluxTransformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+                Input `hidden_states`.
+            controlnet_cond (`torch.Tensor`):
+                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
+            encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
+                Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
+            pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
+                from the embeddings of input conditions.
+            timestep ( `torch.LongTensor`):
+                Used to indicate denoising step.
+            block_controlnet_hidden_states: (`list` of `torch.Tensor`):
+                A list of tensors that if specified are added to the residuals of transformer blocks.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        if joint_attention_kwargs is not None:
+            joint_attention_kwargs = joint_attention_kwargs.copy()
+            lora_scale = joint_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+        hidden_states = self.img_in(hidden_states)
+
+        # add
+        hidden_states = hidden_states + self.controlnet_x_embedder(controlnet_cond)
+
+        temb = self.time_text_embed(timestep, hidden_states)
+
+        image_rotary_emb = self.pos_embed(img_shapes, txt_seq_lens, device=hidden_states.device)
+
+        timestep = timestep.to(hidden_states.dtype)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        block_samples = ()
+        for index_block, block in enumerate(self.transformer_blocks):
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+                encoder_hidden_states, hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    encoder_hidden_states_mask,
+                    temb,
+                    image_rotary_emb,
+                )
+
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                )
+            block_samples = block_samples + (hidden_states,)
+
+        # controlnet block
+        controlnet_block_samples = ()
+        for block_sample, controlnet_block in zip(block_samples, self.controlnet_blocks):
+            block_sample = controlnet_block(block_sample)
+            controlnet_block_samples = controlnet_block_samples + (block_sample,)
+
+        # scaling
+        controlnet_block_samples = [sample * conditioning_scale for sample in controlnet_block_samples]
+        controlnet_block_samples = None if len(controlnet_block_samples) == 0 else controlnet_block_samples
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return controlnet_block_samples
+
+        return QwenImageControlNetOutput(
+            controlnet_block_samples=controlnet_block_samples,
+        )
+
+
+class QwenImageMultiControlNetModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    r"""
+    `QwenImageMultiControlNetModel` wrapper class for Multi-QwenImageControlNetModel
+
+    This module is a wrapper for multiple instances of the `QwenImageControlNetModel`. The `forward()` API is designed
+    to be compatible with `QwenImageControlNetModel`.
+
+    Args:
+        controlnets (`List[QwenImageControlNetModel]`):
+            Provides additional conditioning to the unet during the denoising process. You must set multiple
+            `QwenImageControlNetModel` as a list.
+    """
+
+    def __init__(self, controlnets):
+        super().__init__()
+        self.nets = nn.ModuleList(controlnets)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        controlnet_cond: List[torch.tensor],
+        conditioning_scale: List[float],
+        encoder_hidden_states: torch.Tensor = None,
+        encoder_hidden_states_mask: torch.Tensor = None,
+        timestep: torch.LongTensor = None,
+        img_shapes: Optional[List[Tuple[int, int, int]]] = None,
+        txt_seq_lens: Optional[List[int]] = None,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[QwenImageControlNetOutput, Tuple]:
+        # ControlNet-Union with multiple conditions
+        # only load one ControlNet for saving memories
+        if len(self.nets) == 1:
+            controlnet = self.nets[0]
+
+            for i, (image, scale) in enumerate(zip(controlnet_cond, conditioning_scale)):
+                block_samples = controlnet(
+                    hidden_states=hidden_states,
+                    controlnet_cond=image,
+                    conditioning_scale=scale,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    timestep=timestep,
+                    img_shapes=img_shapes,
+                    txt_seq_lens=txt_seq_lens,
+                    joint_attention_kwargs=joint_attention_kwargs,
+                    return_dict=return_dict,
+                )
+
+                # merge samples
+                if i == 0:
+                    control_block_samples = block_samples
+                else:
+                    if block_samples is not None and control_block_samples is not None:
+                        control_block_samples = [
+                            control_block_sample + block_sample
+                            for control_block_sample, block_sample in zip(control_block_samples, block_samples)
+                        ]
+        else:
+            raise ValueError("QwenImageMultiControlNetModel only supports a single controlnet-union now.")
+
+        return control_block_samples

src/diffusers/models/transformers/transformer_qwenimage.py

@@ -16,6 +16,7 @@ import functools
 import math
 from typing import Any, Dict, List, Optional, Tuple, Union
 
+import numpy as np
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
@@ -552,6 +553,7 @@ class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fro
         txt_seq_lens: Optional[List[int]] = None,
         guidance: torch.Tensor = None,  # TODO: this should probably be removed
         attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_block_samples=None,
         return_dict: bool = True,
     ) -> Union[torch.Tensor, Transformer2DModelOutput]:
         """
@@ -631,6 +633,12 @@ class QwenImageTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fro
                     joint_attention_kwargs=attention_kwargs,
                 )
 
+            # controlnet residual
+            if controlnet_block_samples is not None:
+                interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
+                interval_control = int(np.ceil(interval_control))
+                hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
+
         # Use only the image part (hidden_states) from the dual-stream blocks
         hidden_states = self.norm_out(hidden_states, temb)
         output = self.proj_out(hidden_states)

src/diffusers/modular_pipelines/modular_pipeline_utils.py

@@ -209,7 +209,7 @@ class ComponentSpec:
 
         # Get all loading fields in order
         loading_fields = cls.loading_fields()
-        result = {f: None for f in loading_fields}
+        result = dict.fromkeys(loading_fields)
 
         if load_id == "null":
             return result

src/diffusers/pipelines/__init__.py

@@ -393,6 +393,7 @@ else:
         "QwenImageImg2ImgPipeline",
         "QwenImageInpaintPipeline",
         "QwenImageEditPipeline",
+        "QwenImageControlNetPipeline",
     ]
 try:
     if not is_onnx_available():
@@ -712,6 +713,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .pia import PIAPipeline
         from .pixart_alpha import PixArtAlphaPipeline, PixArtSigmaPipeline
         from .qwenimage import (
+            QwenImageControlNetPipeline,
             QwenImageEditPipeline,
             QwenImageImg2ImgPipeline,
             QwenImageInpaintPipeline,

src/diffusers/pipelines/qwenimage/__init__.py

@@ -24,6 +24,7 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["modeling_qwenimage"] = ["ReduxImageEncoder"]
     _import_structure["pipeline_qwenimage"] = ["QwenImagePipeline"]
+    _import_structure["pipeline_qwenimage_controlnet"] = ["QwenImageControlNetPipeline"]
     _import_structure["pipeline_qwenimage_edit"] = ["QwenImageEditPipeline"]
     _import_structure["pipeline_qwenimage_img2img"] = ["QwenImageImg2ImgPipeline"]
     _import_structure["pipeline_qwenimage_inpaint"] = ["QwenImageInpaintPipeline"]
@@ -36,6 +37,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from ...utils.dummy_torch_and_transformers_objects import *  # noqa F403
     else:
         from .pipeline_qwenimage import QwenImagePipeline
+        from .pipeline_qwenimage_controlnet import QwenImageControlNetPipeline
         from .pipeline_qwenimage_edit import QwenImageEditPipeline
         from .pipeline_qwenimage_img2img import QwenImageImg2ImgPipeline
         from .pipeline_qwenimage_inpaint import QwenImageInpaintPipeline

src/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet.py

@@ -0,0 +1,948 @@
+# Copyright 2025 Qwen-Image Team, InstantX Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import QwenImageLoraLoaderMixin
+from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel
+from ...models.controlnets.controlnet_qwenimage import QwenImageControlNetModel, QwenImageMultiControlNetModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import QwenImagePipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import load_image
+        >>> from diffusers import QwenImageControlNetModel, QwenImageMultiControlNetModel, QwenImageControlNetPipeline
+
+        >>> # QwenImageControlNetModel
+        >>> controlnet = QwenImageControlNetModel.from_pretrained(
+        ...     "InstantX/Qwen-Image-ControlNet-Union", torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe = QwenImageControlNetPipeline.from_pretrained(
+        ...     "Qwen/Qwen-Image", controlnet=controlnet, torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+        >>> prompt = "Aesthetics art, traditional asian pagoda, elaborate golden accents, sky blue and white color palette, swirling cloud pattern, digital illustration, east asian architecture, ornamental rooftop, intricate detailing on building, cultural representation."
+        >>> negative_prompt = " "
+        >>> control_image = load_image(
+        ...     "https://huggingface.co/InstantX/Qwen-Image-ControlNet-Union/resolve/main/conds/canny.png"
+        ... )
+        >>> # Depending on the variant being used, the pipeline call will slightly vary.
+        >>> # Refer to the pipeline documentation for more details.
+        >>> image = pipe(
+        ...     prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     control_image=control_image,
+        ...     controlnet_conditioning_scale=1.0,
+        ...     num_inference_steps=30,
+        ...     true_cfg_scale=4.0,
+        ... ).images[0]
+        >>> image.save("qwenimage_cn_union.png")
+
+        >>> # QwenImageMultiControlNetModel
+        >>> controlnet = QwenImageControlNetModel.from_pretrained(
+        ...     "InstantX/Qwen-Image-ControlNet-Union", torch_dtype=torch.bfloat16
+        ... )
+        >>> controlnet = QwenImageMultiControlNetModel([controlnet])
+        >>> pipe = QwenImageControlNetPipeline.from_pretrained(
+        ...     "Qwen/Qwen-Image", controlnet=controlnet, torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe.to("cuda")
+        >>> prompt = "Aesthetics art, traditional asian pagoda, elaborate golden accents, sky blue and white color palette, swirling cloud pattern, digital illustration, east asian architecture, ornamental rooftop, intricate detailing on building, cultural representation."
+        >>> negative_prompt = " "
+        >>> control_image = load_image(
+        ...     "https://huggingface.co/InstantX/Qwen-Image-ControlNet-Union/resolve/main/conds/canny.png"
+        ... )
+        >>> # Depending on the variant being used, the pipeline call will slightly vary.
+        >>> # Refer to the pipeline documentation for more details.
+        >>> image = pipe(
+        ...     prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     control_image=[control_image, control_image],
+        ...     controlnet_conditioning_scale=[0.5, 0.5],
+        ...     num_inference_steps=30,
+        ...     true_cfg_scale=4.0,
+        ... ).images[0]
+        >>> image.save("qwenimage_cn_union_multi.png")
+        ```
+"""
+
+
+# Coped from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class QwenImageControlNetPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin):
+    r"""
+    The QwenImage pipeline for text-to-image generation.
+
+    Args:
+        transformer ([`QwenImageTransformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`Qwen2.5-VL-7B-Instruct`]):
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the
+            [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant.
+        tokenizer (`QwenTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLQwenImage,
+        text_encoder: Qwen2_5_VLForConditionalGeneration,
+        tokenizer: Qwen2Tokenizer,
+        transformer: QwenImageTransformer2DModel,
+        controlnet: Union[QwenImageControlNetModel, QwenImageMultiControlNetModel],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+            controlnet=controlnet,
+        )
+        self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
+        # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
+        # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
+        self.tokenizer_max_length = 1024
+        self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.prompt_template_encode_start_idx = 34
+        self.default_sample_size = 128
+
+    # Coped from diffusers.pipelines.qwenimage.pipeline_qwenimage.extract_masked_hidden
+    def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor):
+        bool_mask = mask.bool()
+        valid_lengths = bool_mask.sum(dim=1)
+        selected = hidden_states[bool_mask]
+        split_result = torch.split(selected, valid_lengths.tolist(), dim=0)
+
+        return split_result
+
+    # Coped from diffusers.pipelines.qwenimage.pipeline_qwenimage.get_qwen_prompt_embeds
+    def _get_qwen_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        template = self.prompt_template_encode
+        drop_idx = self.prompt_template_encode_start_idx
+        txt = [template.format(e) for e in prompt]
+        txt_tokens = self.tokenizer(
+            txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt"
+        ).to(self.device)
+        encoder_hidden_states = self.text_encoder(
+            input_ids=txt_tokens.input_ids,
+            attention_mask=txt_tokens.attention_mask,
+            output_hidden_states=True,
+        )
+        hidden_states = encoder_hidden_states.hidden_states[-1]
+        split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask)
+        split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
+        attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
+        max_seq_len = max([e.size(0) for e in split_hidden_states])
+        prompt_embeds = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states]
+        )
+        encoder_attention_mask = torch.stack(
+            [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list]
+        )
+
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        return prompt_embeds, encoder_attention_mask
+
+    # Coped from diffusers.pipelines.qwenimage.pipeline_qwenimage.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 1024,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device)
+
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+        prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len)
+
+        return prompt_embeds, prompt_embeds_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_embeds_mask=None,
+        negative_prompt_embeds_mask=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
+            logger.warning(
+                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and prompt_embeds_mask is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `prompt_embeds_mask` also have to be passed. Make sure to generate `prompt_embeds_mask` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+        if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_mask` also have to be passed. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        if max_sequence_length is not None and max_sequence_length > 1024:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}")
+
+    @staticmethod
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents
+    def _pack_latents(latents, batch_size, num_channels_latents, height, width):
+        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
+        latents = latents.permute(0, 2, 4, 1, 3, 5)
+        latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)
+
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents
+    def _unpack_latents(latents, height, width, vae_scale_factor):
+        batch_size, num_patches, channels = latents.shape
+
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (vae_scale_factor * 2))
+        width = 2 * (int(width) // (vae_scale_factor * 2))
+
+        latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2)
+        latents = latents.permute(0, 3, 1, 4, 2, 5)
+
+        latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width)
+
+        return latents
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.prepare_latents
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
+
+        shape = (batch_size, 1, num_channels_latents, height, width)
+
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
+
+        return latents
+
+    # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+        guess_mode=False,
+    ):
+        if isinstance(image, torch.Tensor):
+            pass
+        else:
+            image = self.image_processor.preprocess(image, height=height, width=width)
+
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        if do_classifier_free_guidance and not guess_mode:
+            image = torch.cat([image] * 2)
+
+        return image
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        true_cfg_scale: float = 4.0,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        sigmas: Optional[List[float]] = None,
+        guidance_scale: float = 1.0,
+        control_guidance_start: Union[float, List[float]] = 0.0,
+        control_guidance_end: Union[float, List[float]] = 1.0,
+        control_image: PipelineImageInput = None,
+        controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+        num_images_per_prompt: int = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is
+                not greater than `1`).
+            true_cfg_scale (`float`, *optional*, defaults to 1.0):
+                When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, *optional*, defaults to 3.5):
+                Guidance scale as defined in [Classifier-Free Diffusion
+                Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2.
+                of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting
+                `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to
+                the text `prompt`, usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will be generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`:
+            [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
+            returning a tuple, the first element is a list with the generated images.
+        """
+
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
+
+        if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+            control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+        elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+            control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+        elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+            mult = len(control_image) if isinstance(self.controlnet, QwenImageMultiControlNetModel) else 1
+            control_guidance_start, control_guidance_end = (
+                mult * [control_guidance_start],
+                mult * [control_guidance_end],
+            )
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            negative_prompt_embeds_mask=negative_prompt_embeds_mask,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        has_neg_prompt = negative_prompt is not None or (
+            negative_prompt_embeds is not None and negative_prompt_embeds_mask is not None
+        )
+        do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
+        prompt_embeds, prompt_embeds_mask = self.encode_prompt(
+            prompt=prompt,
+            prompt_embeds=prompt_embeds,
+            prompt_embeds_mask=prompt_embeds_mask,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+        )
+        if do_true_cfg:
+            negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
+                prompt=negative_prompt,
+                prompt_embeds=negative_prompt_embeds,
+                prompt_embeds_mask=negative_prompt_embeds_mask,
+                device=device,
+                num_images_per_prompt=num_images_per_prompt,
+                max_sequence_length=max_sequence_length,
+            )
+
+        # 3. Prepare control image
+        num_channels_latents = self.transformer.config.in_channels // 4
+        if isinstance(self.controlnet, QwenImageControlNetModel):
+            control_image = self.prepare_image(
+                image=control_image,
+                width=width,
+                height=height,
+                batch_size=batch_size * num_images_per_prompt,
+                num_images_per_prompt=num_images_per_prompt,
+                device=device,
+                dtype=self.vae.dtype,
+            )
+            height, width = control_image.shape[-2:]
+
+            if control_image.ndim == 4:
+                control_image = control_image.unsqueeze(2)
+
+            # vae encode
+            self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample)
+            latents_mean = (torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1)).to(
+                device
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                device
+            )
+
+            control_image = retrieve_latents(self.vae.encode(control_image), generator=generator)
+            control_image = (control_image - latents_mean) * latents_std
+
+            control_image = control_image.permute(0, 2, 1, 3, 4)
+
+            # pack
+            control_image = self._pack_latents(
+                control_image,
+                batch_size=control_image.shape[0],
+                num_channels_latents=num_channels_latents,
+                height=control_image.shape[3],
+                width=control_image.shape[4],
+            ).to(dtype=prompt_embeds.dtype, device=device)
+
+        else:
+            if isinstance(self.controlnet, QwenImageMultiControlNetModel):
+                control_images = []
+                for control_image_ in control_image:
+                    control_image_ = self.prepare_image(
+                        image=control_image_,
+                        width=width,
+                        height=height,
+                        batch_size=batch_size * num_images_per_prompt,
+                        num_images_per_prompt=num_images_per_prompt,
+                        device=device,
+                        dtype=self.vae.dtype,
+                    )
+
+                    height, width = control_image_.shape[-2:]
+
+                    if control_image_.ndim == 4:
+                        control_image_ = control_image_.unsqueeze(2)
+
+                    # vae encode
+                    self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample)
+                    latents_mean = (
+                        torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1)
+                    ).to(device)
+                    latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(
+                        1, self.vae.config.z_dim, 1, 1, 1
+                    ).to(device)
+
+                    control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator)
+                    control_image_ = (control_image_ - latents_mean) * latents_std
+
+                    control_image_ = control_image_.permute(0, 2, 1, 3, 4)
+
+                    # pack
+                    control_image_ = self._pack_latents(
+                        control_image_,
+                        batch_size=control_image_.shape[0],
+                        num_channels_latents=num_channels_latents,
+                        height=control_image_.shape[3],
+                        width=control_image_.shape[4],
+                    ).to(dtype=prompt_embeds.dtype, device=device)
+
+                    control_images.append(control_image_)
+
+                control_image = control_images
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+        img_shapes = [(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)] * batch_size
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
+        image_seq_len = latents.shape[1]
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.15),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        controlnet_keep = []
+        for i in range(len(timesteps)):
+            keeps = [
+                1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+                for s, e in zip(control_guidance_start, control_guidance_end)
+            ]
+            controlnet_keep.append(keeps[0] if isinstance(self.controlnet, QwenImageControlNetModel) else keeps)
+
+        # handle guidance
+        if self.transformer.config.guidance_embeds:
+            guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
+            guidance = guidance.expand(latents.shape[0])
+        else:
+            guidance = None
+
+        if self.attention_kwargs is None:
+            self._attention_kwargs = {}
+
+        # 6. Denoising loop
+        self.scheduler.set_begin_index(0)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                if isinstance(controlnet_keep[i], list):
+                    cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+                else:
+                    controlnet_cond_scale = controlnet_conditioning_scale
+                    if isinstance(controlnet_cond_scale, list):
+                        controlnet_cond_scale = controlnet_cond_scale[0]
+                    cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+                # controlnet
+                controlnet_block_samples = self.controlnet(
+                    hidden_states=latents,
+                    controlnet_cond=control_image,
+                    conditioning_scale=cond_scale,
+                    timestep=timestep / 1000,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_hidden_states_mask=prompt_embeds_mask,
+                    img_shapes=img_shapes,
+                    txt_seq_lens=prompt_embeds_mask.sum(dim=1).tolist(),
+                    return_dict=False,
+                )
+
+                with self.transformer.cache_context("cond"):
+                    noise_pred = self.transformer(
+                        hidden_states=latents,
+                        timestep=timestep / 1000,
+                        encoder_hidden_states=prompt_embeds,
+                        encoder_hidden_states_mask=prompt_embeds_mask,
+                        img_shapes=img_shapes,
+                        txt_seq_lens=prompt_embeds_mask.sum(dim=1).tolist(),
+                        controlnet_block_samples=controlnet_block_samples,
+                        attention_kwargs=self.attention_kwargs,
+                        return_dict=False,
+                    )[0]
+
+                if do_true_cfg:
+                    with self.transformer.cache_context("uncond"):
+                        neg_noise_pred = self.transformer(
+                            hidden_states=latents,
+                            timestep=timestep / 1000,
+                            guidance=guidance,
+                            encoder_hidden_states_mask=negative_prompt_embeds_mask,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            img_shapes=img_shapes,
+                            txt_seq_lens=negative_prompt_embeds_mask.sum(dim=1).tolist(),
+                            controlnet_block_samples=controlnet_block_samples,
+                            attention_kwargs=self.attention_kwargs,
+                            return_dict=False,
+                        )[0]
+                    comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
+
+                    cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
+                    noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True)
+                    noise_pred = comb_pred * (cond_norm / noise_norm)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+        if output_type == "latent":
+            image = latents
+        else:
+            latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
+            latents = latents.to(self.vae.dtype)
+            latents_mean = (
+                torch.tensor(self.vae.config.latents_mean)
+                .view(1, self.vae.config.z_dim, 1, 1, 1)
+                .to(latents.device, latents.dtype)
+            )
+            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
+                latents.device, latents.dtype
+            )
+            latents = latents / latents_std + latents_mean
+            image = self.vae.decode(latents, return_dict=False)[0][:, :, 0]
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image,)
+
+        return QwenImagePipelineOutput(images=image)

src/diffusers/utils/dummy_pt_objects.py

@@ -1083,6 +1083,36 @@ class PriorTransformer(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class QwenImageControlNetModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
+class QwenImageMultiControlNetModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class QwenImageTransformer2DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1757,6 +1757,21 @@ class PixArtSigmaPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class QwenImageControlNetPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class QwenImageEditPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]